1
00:00:00,000 --> 00:00:04,179
If I say to you, Mary likes her, you instantly

2
00:00:04,179 --> 00:00:06,280
know I'm talking about two totally different

3
00:00:06,280 --> 00:00:09,400
people, like Mary likes Susan, or Mary likes

4
00:00:09,400 --> 00:00:12,880
Jane. Yeah, two distinct entities. But if I just

5
00:00:12,880 --> 00:00:15,599
add a few letters and say, Mary likes herself,

6
00:00:15,820 --> 00:00:17,760
you immediately know it's the exact same person.

7
00:00:17,960 --> 00:00:21,570
Instantly. Mary likes Mary. And it feels incredibly

8
00:00:21,570 --> 00:00:24,070
simple, right? I mean, it feels like the kind

9
00:00:24,070 --> 00:00:26,210
of thing you just know without even really thinking

10
00:00:26,210 --> 00:00:29,329
about it. Oh, absolutely. But, well... What if

11
00:00:29,329 --> 00:00:31,449
I told you that in the literal milliseconds it

12
00:00:31,449 --> 00:00:33,409
took for you to understand the difference between

13
00:00:33,409 --> 00:00:37,469
her and herself, your brain actually executed

14
00:00:37,469 --> 00:00:40,149
this, like, invisible piece of lightning -fast

15
00:00:40,149 --> 00:00:43,090
algebra. It really is like algebra. Yeah. A whole

16
00:00:43,090 --> 00:00:45,509
series of complex mathematical constraints that,

17
00:00:45,770 --> 00:00:47,450
you know, you were never taught in school, but

18
00:00:47,450 --> 00:00:49,869
that you follow flawlessly every single time

19
00:00:49,869 --> 00:00:52,409
you open your mouth. It is a stunning realization

20
00:00:52,409 --> 00:00:55,369
when you first encounter it. Because we walk

21
00:00:55,369 --> 00:00:58,270
around all day. generating and decoding just

22
00:00:58,270 --> 00:01:00,530
massive strings of data. All day long. Right.

23
00:01:00,609 --> 00:01:02,929
And we're completely blind to the sophisticated

24
00:01:02,929 --> 00:01:04,890
machinery operating just beneath the surface

25
00:01:04,890 --> 00:01:07,430
of our own minds. Totally blind to it. When we

26
00:01:07,430 --> 00:01:10,450
casually use a pronoun, we aren't just picking

27
00:01:10,450 --> 00:01:13,609
words from a list. We are navigating a deeply

28
00:01:13,609 --> 00:01:17,280
structured biological syntax. And mapping that

29
00:01:17,280 --> 00:01:19,680
hidden machinery is exactly what we're doing

30
00:01:19,680 --> 00:01:22,519
today on this Deep Dive. Yes, we are. We are

31
00:01:22,519 --> 00:01:25,000
pulling apart a truly seminal textbook in the

32
00:01:25,000 --> 00:01:27,719
field of linguistics. It's called Syntactic Theory,

33
00:01:27,859 --> 00:01:30,420
a Formal Introduction, the second edition. By

34
00:01:30,420 --> 00:01:33,560
Ivan A. Sag, Thomas Wasow, and Emily M. Bender.

35
00:01:33,920 --> 00:01:37,159
Right. And we're going to dive really deep into

36
00:01:37,159 --> 00:01:39,459
the source code of human language today. That's

37
00:01:39,459 --> 00:01:42,040
a good way to put it, the source code. We want

38
00:01:42,040 --> 00:01:44,379
to understand the exact mechanics of how humans

39
00:01:44,379 --> 00:01:47,620
combine words into meaningful structures, because

40
00:01:47,620 --> 00:01:49,879
it's wild once you get into it. Which I should

41
00:01:49,879 --> 00:01:52,959
add is an immensely ambitious undertaking. Oh,

42
00:01:52,959 --> 00:01:56,359
for sure. I mean, we aren't here to look at grammar

43
00:01:56,359 --> 00:01:58,939
as a set of polite suggestions. No, not at all.

44
00:01:59,140 --> 00:02:02,900
We are here to essentially reverse engineer a

45
00:02:02,900 --> 00:02:06,019
rigorous, systematic body of unconscious knowledge.

46
00:02:06,079 --> 00:02:09,560
Yeah. Because understanding this invisible architecture

47
00:02:10,050 --> 00:02:12,330
is really the key to everything from decoding

48
00:02:12,330 --> 00:02:15,629
human cognitive psychology to, well, actually

49
00:02:15,629 --> 00:02:18,030
building artificial intelligence that can genuinely

50
00:02:18,030 --> 00:02:21,090
understand us. Right. The stakes are huge. But

51
00:02:21,090 --> 00:02:23,469
I feel like before we can even look at the underlying

52
00:02:23,469 --> 00:02:25,969
code, we really have to clear away some serious

53
00:02:25,969 --> 00:02:28,889
mental baggage. Oh, the grammar trauma. Yes.

54
00:02:29,270 --> 00:02:31,610
Exactly. Yes. Because I know that when a lot

55
00:02:31,610 --> 00:02:33,870
of people hear words like syntax or grammar,

56
00:02:34,389 --> 00:02:36,909
this very specific, slightly traumatic image

57
00:02:36,909 --> 00:02:38,949
comes to mind. The middle school English teacher.

58
00:02:39,150 --> 00:02:42,180
Right. You picture this teacher sternly telling

59
00:02:42,180 --> 00:02:44,240
you to never split an infinitive or, you know,

60
00:02:44,379 --> 00:02:46,139
drawing those big red circles on your paper when

61
00:02:46,139 --> 00:02:48,379
you end a sentence with a preposition. Yes, and

62
00:02:48,379 --> 00:02:50,759
that is exactly the baggage the authors of the

63
00:02:50,759 --> 00:02:53,379
textbook discard right at the outset. Thank goodness.

64
00:02:54,099 --> 00:02:56,659
They draw a hard, impenetrable line between what

65
00:02:56,659 --> 00:02:59,840
we call prescriptive grammar and descriptive

66
00:02:59,840 --> 00:03:02,560
grammar. OK, let's unpack that difference. Well,

67
00:03:02,620 --> 00:03:04,800
what you experience in middle school is prescriptive

68
00:03:04,800 --> 00:03:07,819
grammar. It's essentially a set of social admonitions.

69
00:03:07,990 --> 00:03:11,229
Like rules for polite society. Right. It's an

70
00:03:11,229 --> 00:03:13,669
attempt to legislate how language should be used.

71
00:03:14,189 --> 00:03:16,270
And honestly, it's mostly based on historical

72
00:03:16,270 --> 00:03:19,129
precedents from Latin that have absolutely no

73
00:03:19,129 --> 00:03:21,949
bearing on the actual biological reality of English.

74
00:03:22,610 --> 00:03:24,210
I've always kind of thought of prescriptive grammar

75
00:03:24,210 --> 00:03:26,289
like a dress code. Oh, that's a good analogy.

76
00:03:26,530 --> 00:03:28,449
Yeah. It's society deciding that, I don't know,

77
00:03:28,469 --> 00:03:30,409
you have to wear a tie to a fancy restaurant.

78
00:03:30,789 --> 00:03:33,270
It's completely made up. It's just a social construct.

79
00:03:33,610 --> 00:03:36,180
Exactly. Descriptive grammar, on the other hand,

80
00:03:36,300 --> 00:03:39,060
is like the actual anatomy of the human body.

81
00:03:39,340 --> 00:03:41,659
Yes. It's the muscles, the bones, the tendons.

82
00:03:41,879 --> 00:03:44,900
You can't legislate how a knee bends. It just

83
00:03:44,900 --> 00:03:46,919
bends that way naturally because of its underlying

84
00:03:46,919 --> 00:03:50,039
geometry. That is a highly accurate way to frame

85
00:03:50,039 --> 00:03:52,819
it. The textbook authors actually point out that

86
00:03:52,819 --> 00:03:55,580
trying to legislate a naturally occurring cognitive

87
00:03:55,580 --> 00:03:58,719
system, like language, is fundamentally absurd.

88
00:03:59,439 --> 00:04:02,259
Because it's biology. Right. They highlight this

89
00:04:02,259 --> 00:04:06,169
with a famous historical anecdote. In 1897, the

90
00:04:06,169 --> 00:04:08,409
Indiana state legislature actually considered

91
00:04:08,409 --> 00:04:10,830
a bill that would have legally established the

92
00:04:10,830 --> 00:04:14,969
value of pi as exactly 3 .2. Wait, really? They

93
00:04:14,969 --> 00:04:17,329
actually tried to pass a law changing a mathematical

94
00:04:17,329 --> 00:04:20,250
constant? They did. I mean, it passed the state

95
00:04:20,250 --> 00:04:22,490
house before thankfully dying in the Senate.

96
00:04:23,069 --> 00:04:25,990
That is hilarious. But modern linguists view

97
00:04:25,990 --> 00:04:28,410
prescriptive grammar rules with exactly that

98
00:04:28,410 --> 00:04:31,490
same level of amusement. I love that. You simply

99
00:04:31,490 --> 00:04:34,459
cannot pass a law saying A native English speaker

100
00:04:34,459 --> 00:04:37,579
shouldn't end a sentence with a preposition because

101
00:04:37,579 --> 00:04:39,920
native speakers do it naturally, constantly,

102
00:04:40,100 --> 00:04:43,540
and effortlessly. Right. It's just what the anatomy

103
00:04:43,540 --> 00:04:47,939
does. Exactly. Descriptive linguistics just observes

104
00:04:47,939 --> 00:04:51,000
what native speakers actually do and then works

105
00:04:51,000 --> 00:04:53,259
backward to find the mathematical and logical

106
00:04:53,259 --> 00:04:55,860
rules that naturally govern that behavior. Okay.

107
00:04:55,860 --> 00:04:58,660
So if we aren't studying the dress code and we

108
00:04:58,660 --> 00:05:01,379
are studying the anatomy, I kinda need proof

109
00:05:01,379 --> 00:05:04,120
that this anatomy actually exists. Fair enough.

110
00:05:04,600 --> 00:05:06,800
Because, frankly, I don't feel like I'm doing

111
00:05:06,800 --> 00:05:09,220
algebra when I talk. I just sort of open my mouth

112
00:05:09,220 --> 00:05:11,420
and the words come out. Most people feel that

113
00:05:11,420 --> 00:05:13,600
way. So how do we actually know there are hidden

114
00:05:13,600 --> 00:05:16,019
rules? Let's go back to the puzzle you introduced

115
00:05:16,019 --> 00:05:18,860
at the very beginning. The reflexive pronoun

116
00:05:18,860 --> 00:05:21,939
puzzle. Okay. Words like herself, ourselves,

117
00:05:22,360 --> 00:05:25,620
myself. Right. Let's try to figure out the actual

118
00:05:25,620 --> 00:05:28,240
unconscious rule your brain uses for these words.

119
00:05:28,550 --> 00:05:31,670
We know she likes her means two different people

120
00:05:31,670 --> 00:05:36,029
and she likes herself means one person. Yep So

121
00:05:36,029 --> 00:05:39,329
as a first pass, how would you define the rule

122
00:05:39,329 --> 00:05:42,350
for when you are allowed to use a reflexive pronoun?

123
00:05:42,910 --> 00:05:45,790
All right. Let's play along. Let's call it hypothesis

124
00:05:45,790 --> 00:05:51,069
one go for it. I'd say a Reflexive pronoun can

125
00:05:51,069 --> 00:05:53,709
only be used if there is a word earlier in the

126
00:05:53,709 --> 00:05:55,810
sentence that refers to the exact same person

127
00:05:55,810 --> 00:06:00,060
Okay, so in a sentence like She voted for herself.

128
00:06:00,519 --> 00:06:03,339
The word she is right there at the start. It

129
00:06:03,339 --> 00:06:06,860
matches herself. The rule works perfectly. It

130
00:06:06,860 --> 00:06:09,199
seems like a perfectly logical, straightforward

131
00:06:09,199 --> 00:06:13,000
rule. Thanks. But human language is rarely that

132
00:06:13,000 --> 00:06:16,519
simple. Let's stress test hypothesis I with a

133
00:06:16,519 --> 00:06:18,839
slightly more complex sentence. Bring it on.

134
00:06:19,220 --> 00:06:22,259
A sentence with a subordinate clause. How about,

135
00:06:22,560 --> 00:06:26,000
we think that Leslie likes us. We think that

136
00:06:26,000 --> 00:06:28,889
Leslie likes us. Yeah, that sounds totally normal

137
00:06:28,889 --> 00:06:33,490
to me. Now, apply your rule. The word we is at

138
00:06:33,490 --> 00:06:36,009
the beginning of the sentence, so according to

139
00:06:36,009 --> 00:06:38,850
Hypothesis I, you should be perfectly allowed

140
00:06:38,850 --> 00:06:41,089
to use a reflexive pronoun later in the sentence

141
00:06:41,089 --> 00:06:46,810
that refers back to we. Try it. Okay, so we think

142
00:06:46,810 --> 00:06:50,240
that Leslie likes ourselves. How does that sound?

143
00:06:51,180 --> 00:06:54,550
Oh. No, that sounds completely wrong. It does.

144
00:06:54,920 --> 00:06:57,500
Like, it actually crashes my brain a little bit

145
00:06:57,500 --> 00:07:00,120
to say it out loud. We think that Leslie likes

146
00:07:00,120 --> 00:07:04,240
ourselves. Yeah, no. Exactly. Every native English

147
00:07:04,240 --> 00:07:06,800
speaker instantly rejects that sentence. Wow.

148
00:07:07,079 --> 00:07:09,420
But Hypothesis I says it should be perfectly

149
00:07:09,420 --> 00:07:12,860
fine, because we and ourselves refer to the same

150
00:07:12,860 --> 00:07:15,819
group of people, and we comes earlier in the

151
00:07:15,819 --> 00:07:20,300
sentence. Right. So Hypothesis I is dead. The

152
00:07:20,300 --> 00:07:22,879
anatomy is clearly more complex than just scanning

153
00:07:22,879 --> 00:07:25,500
backward for a matching word. Okay, okay, let

154
00:07:25,500 --> 00:07:28,000
me revise it then. Let's build hypothesis 2.

155
00:07:28,360 --> 00:07:31,420
Let's hear it. The problem there was that we

156
00:07:31,420 --> 00:07:35,060
and ourselves were in two totally different parts

157
00:07:35,060 --> 00:07:38,120
of the sentence. Right. We think is one part

158
00:07:38,120 --> 00:07:41,379
and Leslie likes ourselves is another part. They're

159
00:07:41,379 --> 00:07:44,319
entirely different clauses. Yes, they are. So,

160
00:07:44,480 --> 00:07:48,360
hypothesis 2. A reflexive pronoun can only appear

161
00:07:48,360 --> 00:07:51,060
if there is a matching word in the exact same

162
00:07:51,060 --> 00:07:54,680
clause. A very solid scientific adjustment. You've

163
00:07:54,680 --> 00:07:57,279
localized the constraint. Makes sense, right?

164
00:07:57,860 --> 00:08:01,800
It does. But let's test hypothesis two. Consider

165
00:08:01,800 --> 00:08:05,360
the phrase, those pictures of us offended us.

166
00:08:06,050 --> 00:08:09,410
Those pictures of us offended us. That's a single

167
00:08:09,410 --> 00:08:12,550
clause. It sounds fine. Now swap the final pronoun

168
00:08:12,550 --> 00:08:15,990
for reflexive. Those pictures of us offended

169
00:08:15,990 --> 00:08:19,879
ourselves. Those pictures of us offended ourselves.

170
00:08:21,009 --> 00:08:23,790
Wow. Awkward, right? That sounds even worse than

171
00:08:23,790 --> 00:08:25,850
the last one. It sounds like a glitching robot

172
00:08:25,850 --> 00:08:28,509
trying to speak English. It really does. But

173
00:08:28,509 --> 00:08:31,170
wait, us and ourselves are in the exact same

174
00:08:31,170 --> 00:08:34,309
clause there, so Hypothesis 2 totally fails too.

175
00:08:34,649 --> 00:08:38,110
It does. I don't get it. What is the actual rule

176
00:08:38,110 --> 00:08:41,250
then? Well, what's fascinating here is that to

177
00:08:41,250 --> 00:08:44,490
find the real rule, Hypothesis 3, we have to

178
00:08:44,490 --> 00:08:46,990
stop looking at sentences as just flat strings

179
00:08:46,990 --> 00:08:49,710
of words. Okay. We have to start looking at the

180
00:08:49,710 --> 00:08:52,840
underlying geometry. The textbook introduces

181
00:08:52,840 --> 00:08:57,639
this concept of arguments. Arguments? Like fighting?

182
00:08:57,960 --> 00:09:01,740
No, no. Think of a verb like offend or like as

183
00:09:01,740 --> 00:09:04,659
a central action. Okay. The arguments are the

184
00:09:04,659 --> 00:09:07,379
core participants in that action. The main subject

185
00:09:07,379 --> 00:09:10,279
and the main object. Got it. The main players.

186
00:09:10,539 --> 00:09:13,519
Exactly. The actual hidden rule your brain uses

187
00:09:13,519 --> 00:09:16,779
is this. A reflexive pronoun must be an argument

188
00:09:16,779 --> 00:09:19,360
of a verb that has another preceding argument

189
00:09:19,360 --> 00:09:21,639
with the same referent. Wait, run that by me

190
00:09:21,639 --> 00:09:24,799
again. They must be co -arguments. Both words

191
00:09:24,799 --> 00:09:27,539
have to be direct, core participants of the exact

192
00:09:27,539 --> 00:09:30,799
same verb. I think I need to map that out. Let's

193
00:09:30,799 --> 00:09:32,820
look at that sentence again. Those pictures of

194
00:09:32,820 --> 00:09:35,840
us offended us. Okay, break it down. So the core

195
00:09:35,840 --> 00:09:39,080
action is offended. Who or what is actually doing

196
00:09:39,080 --> 00:09:41,539
the offending? Right. What's the subject? It's

197
00:09:41,539 --> 00:09:44,240
not us. The thing doing the offending is the

198
00:09:44,240 --> 00:09:47,179
entire massive noun phrase, those pictures of

199
00:09:47,179 --> 00:09:49,620
us. That whole chunk is the subject argument.

200
00:09:50,039 --> 00:09:53,509
Precisely. The word us is just buried inside

201
00:09:53,509 --> 00:09:56,389
that larger subject argument. It is merely a

202
00:09:56,389 --> 00:10:00,210
modifier. Oh. It is not an argument of the verb

203
00:10:00,210 --> 00:10:05,120
offend itself. So us and the final object ourselves

204
00:10:05,120 --> 00:10:08,080
are not co -arguments. They are not. And because

205
00:10:08,080 --> 00:10:10,519
they aren't co -arguments, the rule isn't satisfied

206
00:10:10,519 --> 00:10:12,840
and my brain just immediately flags the sentence

207
00:10:12,840 --> 00:10:16,340
as ungrammatical. Exactly. You reject it without

208
00:10:16,340 --> 00:10:20,389
even knowing why. That is wild. I mean, I have

209
00:10:20,389 --> 00:10:22,549
never once in my life consciously thought about

210
00:10:22,549 --> 00:10:24,549
the concept of a co -argument. I didn't even

211
00:10:24,549 --> 00:10:26,750
know what an argument was until five minutes

212
00:10:26,750 --> 00:10:31,169
ago. Few people do. But my brain was mathematically

213
00:10:31,169 --> 00:10:33,750
enforcing that exact rule when I was a toddler.

214
00:10:34,409 --> 00:10:37,240
And it goes even deeper than that. Seriously?

215
00:10:37,700 --> 00:10:40,899
Oh, yes. Think about an imperative sentence,

216
00:10:41,379 --> 00:10:45,289
a command. Okay. If I look right at you and yell,

217
00:10:45,789 --> 00:10:50,190
vote for yourself. Wait a second. If the rule

218
00:10:50,190 --> 00:10:53,049
requires two co -arguments, a subject and an

219
00:10:53,049 --> 00:10:56,950
object, vote for yourself shouldn't work. Why

220
00:10:56,950 --> 00:10:59,970
not? Because there is no preceding subject. The

221
00:10:59,970 --> 00:11:02,149
sentence literally starts with the verb, vote.

222
00:11:03,149 --> 00:11:05,330
By all surface logic, the sentence should crash.

223
00:11:05,549 --> 00:11:09,129
Your brain should reject it. But it doesn't.

224
00:11:09,350 --> 00:11:13,080
You understand it perfectly. You do. And the

225
00:11:13,080 --> 00:11:16,100
only way the math holds up is if we concede that

226
00:11:16,100 --> 00:11:19,049
there is a phantom argument. operating in the

227
00:11:19,049 --> 00:11:23,210
syntax. A phantom argument. Yes, there is an

228
00:11:23,210 --> 00:11:26,809
invisible implied you functioning as the subject

229
00:11:26,809 --> 00:11:29,230
argument in the mental architecture of that sentence.

230
00:11:30,029 --> 00:11:34,929
That is wow. Your brain is literally calculating

231
00:11:34,929 --> 00:11:37,669
the syntactic weight of a word that makes no

232
00:11:37,669 --> 00:11:41,190
sound and leaves no ink on the page. It's validating

233
00:11:41,190 --> 00:11:43,450
that phantom word against the reflexive pronoun

234
00:11:43,450 --> 00:11:46,840
and approving the structure. calculating phantom

235
00:11:46,840 --> 00:11:49,559
data points in milliseconds. It really makes

236
00:11:49,559 --> 00:11:52,019
you realize that language isn't just like a cultural

237
00:11:52,019 --> 00:11:54,440
habit we picked up. Not at all. It's a deeply

238
00:11:54,440 --> 00:11:57,519
embedded biological function. Yes it is. Which

239
00:11:57,519 --> 00:11:59,639
honestly makes me wonder, how did linguistics

240
00:11:59,639 --> 00:12:02,379
even figure this out? Because studying hidden

241
00:12:02,379 --> 00:12:05,440
phantom arguments is a long way from just cataloging

242
00:12:05,440 --> 00:12:07,120
different languages and writing dictionaries.

243
00:12:07,419 --> 00:12:10,220
It required a massive paradigm shift in the field.

244
00:12:10,679 --> 00:12:13,100
The textbook actually takes a brief detour to

245
00:12:13,100 --> 00:12:16,360
explain this historical pivot. Oh, right. The

246
00:12:16,360 --> 00:12:19,639
history section. Yeah. For a long time, linguistics

247
00:12:19,639 --> 00:12:22,419
was largely historical. People like Sir William

248
00:12:22,419 --> 00:12:25,220
Jones in the 18th century were figuring out that

249
00:12:25,220 --> 00:12:28,299
Sanskrit, Latin, and Greek all shared a common

250
00:12:28,299 --> 00:12:30,899
ancestor. Mapping out those big family trees

251
00:12:30,899 --> 00:12:33,960
of Indo -European languages. Exactly. Or it was

252
00:12:33,960 --> 00:12:36,500
anthropological, like Franz Boas, studying the

253
00:12:36,500 --> 00:12:39,460
sheer incredible diversity of indigenous American

254
00:12:39,460 --> 00:12:41,559
languages. Just documenting what was out there.

255
00:12:41,720 --> 00:12:45,279
Right. But then... In the mid -20th century,

256
00:12:45,759 --> 00:12:48,340
the behaviorists took over psychology. B .F.

257
00:12:48,500 --> 00:12:50,200
Skinner and his whole framework, right? Yes.

258
00:12:50,679 --> 00:12:52,559
Skinner and the behaviorists argued that the

259
00:12:52,559 --> 00:12:55,279
mind was basically a blank slate. They believed

260
00:12:55,279 --> 00:12:57,720
that language was learned entirely through external

261
00:12:57,720 --> 00:13:00,940
stimulus and response. Like training a pet? Pretty

262
00:13:00,940 --> 00:13:03,600
much. You point to a dog, you say, dog, you get

263
00:13:03,600 --> 00:13:07,039
a reward. Language, in their view, was just a

264
00:13:07,039 --> 00:13:09,460
highly complex habit learned through sheer repetition.

265
00:13:09,659 --> 00:13:12,320
But wait, how do you learn the co -argument rule

266
00:13:12,320 --> 00:13:15,159
for reflexive pronouns through repetition? That's

267
00:13:15,159 --> 00:13:16,940
the million -dollar question. Because nobody

268
00:13:16,940 --> 00:13:18,919
corrects a three -year -old by saying, ah, sweetie,

269
00:13:19,059 --> 00:13:21,539
remember, that pronoun isn't a direct argument

270
00:13:21,539 --> 00:13:23,820
of the verbal action. They certainly don't. They

271
00:13:23,820 --> 00:13:27,179
just naturally absorb it. Which is exactly what

272
00:13:27,179 --> 00:13:30,720
Noam Chomsky pointed out in the late 1950s. Chomsky

273
00:13:30,720 --> 00:13:33,580
wrote this devastating review of Skinner's work,

274
00:13:33,860 --> 00:13:36,399
and it effectively sparked the entire cognitive

275
00:13:36,399 --> 00:13:38,899
revolution. He kind of dismantled the whole theory.

276
00:13:39,000 --> 00:13:42,120
He did. Chomsky looked at the staggering complexity

277
00:13:42,120 --> 00:13:45,080
of syntax, the kind of invisible rules we just

278
00:13:45,080 --> 00:13:48,000
demonstrated, and said that stimulus and response

279
00:13:48,000 --> 00:13:51,519
cannot possibly explain how quickly and flawlessly

280
00:13:51,519 --> 00:13:54,100
children acquire language. The input they get

281
00:13:54,100 --> 00:13:56,940
is just too messy. Exactly. The external input

282
00:13:56,940 --> 00:13:59,879
is fragmented. Therefore, the complexity of language

283
00:13:59,879 --> 00:14:03,460
must be, at least in part, innate. A biological

284
00:14:03,460 --> 00:14:06,659
feature. Yes. He proposed that humans possess

285
00:14:06,659 --> 00:14:10,019
a specialized biological mental organ for language.

286
00:14:10,799 --> 00:14:13,220
Just like a heart or a liver, but for syntax.

287
00:14:13,659 --> 00:14:16,820
Exactly. We are hardwired with a universal grammar,

288
00:14:17,240 --> 00:14:20,360
a basic architectural blueprint for syntax that

289
00:14:20,360 --> 00:14:22,639
sits in our brains before we even hear our first

290
00:14:22,639 --> 00:14:26,019
word. That blows my mind. Whether you end up

291
00:14:26,019 --> 00:14:29,320
speaking English, Tagalog, or Arabic, the underlying

292
00:14:29,320 --> 00:14:33,169
biological constraints are the same. So mapping

293
00:14:33,169 --> 00:14:36,090
syntax is literally mapping the physical parameters

294
00:14:36,090 --> 00:14:39,269
of human thought? That's just incredible. It

295
00:14:39,269 --> 00:14:41,929
really is. But, you know, if this system is so

296
00:14:41,929 --> 00:14:44,070
robust and biological, why is human language

297
00:14:44,070 --> 00:14:47,850
so famously messy? What do you mean? Well, the

298
00:14:47,850 --> 00:14:50,029
textbook spends a lot of time talking about ambiguity.

299
00:14:50,529 --> 00:14:53,950
Uh -huh, yes. Ambiguity. Because ambiguity is

300
00:14:53,950 --> 00:14:56,350
the ultimate stress test for any syntactic theory.

301
00:14:56,470 --> 00:14:59,370
Right. If we want to prove that our mental architecture

302
00:14:59,370 --> 00:15:02,590
exists, we have to look at how it handles massive,

303
00:15:02,990 --> 00:15:05,970
overlapping data in real time. Take the word

304
00:15:05,970 --> 00:15:09,720
run. The text lists an exhausting number of ways

305
00:15:09,720 --> 00:15:11,799
we use it. Oh, that list is amazing. I wrote

306
00:15:11,799 --> 00:15:14,179
some of them down. Cheetahs run down a hill.

307
00:15:14,519 --> 00:15:16,879
The president will run for office. The car won't

308
00:15:16,879 --> 00:15:19,980
run. Right? A trail runs over the hill. Die will

309
00:15:19,980 --> 00:15:22,360
run in the wash. This hotel room will run you

310
00:15:22,360 --> 00:15:25,580
$200. You run a risk. Stockings will run. You

311
00:15:25,580 --> 00:15:28,460
score a run in baseball. It's endless. Over a

312
00:15:28,460 --> 00:15:31,539
dozen completely distinct meanings. And yet,

313
00:15:32,019 --> 00:15:34,519
if I use any of those in a natural conversation

314
00:15:34,519 --> 00:15:36,980
with you, you don't even blink. No, I just get

315
00:15:36,980 --> 00:15:39,740
it. Your brain instantly disambiguates the word

316
00:15:39,740 --> 00:15:42,659
based on its immediate syntactic and semantic

317
00:15:42,659 --> 00:15:45,539
environment. It's seamless. And it's not just

318
00:15:45,539 --> 00:15:48,559
individual words that are ambiguous. Entire structures

319
00:15:48,559 --> 00:15:51,559
are ambiguous. Right. The text uses that classic

320
00:15:51,559 --> 00:15:54,659
example, visiting relatives can be boring. Such

321
00:15:54,659 --> 00:15:57,500
a great example. Is it the act of traveling to

322
00:15:57,500 --> 00:16:00,039
see your relatives that is boring? Or is it the

323
00:16:00,039 --> 00:16:02,179
relatives themselves who have come to visit you

324
00:16:02,179 --> 00:16:04,220
that are boring? You literally can't tell just

325
00:16:04,220 --> 00:16:07,399
from the sound. Exactly. The words are identical.

326
00:16:07,860 --> 00:16:11,299
The pronunciation is identical. The only thing

327
00:16:11,299 --> 00:16:14,299
that differentiates the two meanings is the invisible

328
00:16:14,299 --> 00:16:17,179
syntactic structure your brain overlays onto

329
00:16:17,179 --> 00:16:19,720
those words. My favorite one in the book is,

330
00:16:19,919 --> 00:16:22,840
I saw her duck. Yes. Did I see a woman physically

331
00:16:22,840 --> 00:16:25,440
lower her head to avoid a baseball? Or did I

332
00:16:25,440 --> 00:16:28,019
see the pet waterfowl that belongs to her? And

333
00:16:28,019 --> 00:16:30,820
this brings us to a highly practical point for

334
00:16:30,820 --> 00:16:32,960
you, the listener. Why does all this matter?

335
00:16:33,379 --> 00:16:36,399
Exactly. Why does studying this invisible structure

336
00:16:36,399 --> 00:16:39,179
matter today? Because if you want to build a

337
00:16:39,179 --> 00:16:41,480
search engine, or a machine translation tool,

338
00:16:41,580 --> 00:16:43,980
or a voice assistant that actually understands

339
00:16:43,980 --> 00:16:46,820
human commands... You can't just feed it a dictionary.

340
00:16:47,259 --> 00:16:50,350
No! A computer looking at the string I saw her

341
00:16:50,350 --> 00:16:53,789
duck will completely stall unless it is programmed

342
00:16:53,789 --> 00:16:56,570
with a formal mathematical system to calculate

343
00:16:56,570 --> 00:16:59,950
structural ambiguity. It needs the math. We have

344
00:16:59,950 --> 00:17:02,490
to be able to explicitly write the source code

345
00:17:02,490 --> 00:17:06,089
that our brains just run implicitly. Which brings

346
00:17:06,089 --> 00:17:08,289
us to sort of the core challenge of the textbook,

347
00:17:08,450 --> 00:17:11,170
right? How do we visually and mathematically

348
00:17:11,170 --> 00:17:14,069
represent a sentence? Yes. If we're going to

349
00:17:14,069 --> 00:17:16,509
write the source code, we need a modeling language.

350
00:17:16,779 --> 00:17:19,819
The authors start by examining the simplest possible

351
00:17:19,819 --> 00:17:22,339
theory for modeling language. It's something

352
00:17:22,339 --> 00:17:25,539
called regular expressions or finite state grammar.

353
00:17:26,299 --> 00:17:29,519
OK, regular expressions. If you've ever searched

354
00:17:29,519 --> 00:17:32,859
a massive database or used an advanced find and

355
00:17:32,859 --> 00:17:35,599
replace function in a text editor, you've probably

356
00:17:35,599 --> 00:17:38,359
used regular expressions. Yeah, coders use them

357
00:17:38,359 --> 00:17:41,460
all the time. Right. It's a linear flat sequence

358
00:17:41,460 --> 00:17:45,180
model. It basically says match word A followed

359
00:17:45,180 --> 00:17:48,710
by word B. followed by word C. So it models a

360
00:17:48,710 --> 00:17:50,589
sentence kind of like a string of beads. Look

361
00:17:50,589 --> 00:17:53,529
for a noun, then a verb, then an adjective. Precisely.

362
00:17:53,869 --> 00:17:56,170
But the text aggressively proves that this flat,

363
00:17:56,269 --> 00:17:59,009
linear model completely fails for natural language.

364
00:17:59,609 --> 00:18:02,849
Why does it fail? It seems simple enough. It

365
00:18:02,849 --> 00:18:05,529
fails because human language utilizes infinite

366
00:18:05,529 --> 00:18:09,450
embedding. Recursion. Recursion. You can say,

367
00:18:09,670 --> 00:18:12,470
the mouse ran, or the mouse that the cat chased

368
00:18:12,470 --> 00:18:15,910
ran, or The mouse that the cat that the dog bit

369
00:18:15,910 --> 00:18:19,230
chased ran. Oh, I see. You can just nest phrases

370
00:18:19,230 --> 00:18:21,109
inside of phrases inside of phrases forever.

371
00:18:21,609 --> 00:18:24,849
Exactly. A finite state linear model can't handle

372
00:18:24,849 --> 00:18:28,150
deep nesting. It also fundamentally cannot handle

373
00:18:28,150 --> 00:18:31,230
structural ambiguity because it only sees a flat

374
00:18:31,230 --> 00:18:34,269
line of words. It has no depth. It has no depth

375
00:18:34,269 --> 00:18:36,769
whatsoever. So we need depth. We need a third

376
00:18:36,769 --> 00:18:40,109
dimension. Yes. And that leads us to context

377
00:18:40,109 --> 00:18:43,390
-free phrase structure grammar, or CFG for short.

378
00:18:43,569 --> 00:18:46,289
CFG. This is where the linguists bring out the

379
00:18:46,289 --> 00:18:48,809
chalk and start drawing those big trees on the

380
00:18:48,809 --> 00:18:51,430
blackboard. The famous syntax trees. Right. We

381
00:18:51,430 --> 00:18:54,250
take a whole sentence, we label it S at the top,

382
00:18:54,509 --> 00:18:57,069
and we say it branches downward into two major

383
00:18:57,069 --> 00:19:00,930
chunks, a noun phrase, or NP, and a verb phrase,

384
00:19:01,190 --> 00:19:03,890
VP. And then those branches branch further and

385
00:19:03,890 --> 00:19:07,019
further down to the individual words. CFG was

386
00:19:07,019 --> 00:19:09,599
a monumental leap forward in linguistics, wasn't

387
00:19:09,599 --> 00:19:12,839
it? It really was. It finally allowed us to group

388
00:19:12,839 --> 00:19:15,859
words into hierarchical constituents, unified

389
00:19:15,859 --> 00:19:18,900
blocks of meaning. And the beauty of CFG is that

390
00:19:18,900 --> 00:19:21,240
it elegantly solves that structural ambiguity

391
00:19:21,240 --> 00:19:23,940
problem we were just talking about. Let's visualize

392
00:19:23,940 --> 00:19:27,619
this. The textbook uses the sentence, I saw the

393
00:19:27,619 --> 00:19:30,160
astronomer with a telescope. Another classic

394
00:19:30,160 --> 00:19:33,440
ambiguity. Right. Did I use a telescope to see

395
00:19:33,440 --> 00:19:35,890
him? Or did I look across the room and see an

396
00:19:35,890 --> 00:19:38,250
astronomer who was holding a telescope? If you

397
00:19:38,250 --> 00:19:40,289
just read the words left to right, like a string

398
00:19:40,289 --> 00:19:43,230
of beads, you simply don't know. But if you draw

399
00:19:43,230 --> 00:19:46,779
a CFG tree, the geometry solves it. Walk us through

400
00:19:46,779 --> 00:19:49,680
the two trees. Okay, in my mind's eye, I'm drawing

401
00:19:49,680 --> 00:19:52,460
Tree 1. I have my verb phrase, which contains

402
00:19:52,460 --> 00:19:55,519
the verb saw, and the noun phrase the astronomer.

403
00:19:55,819 --> 00:19:58,400
Okay, you have the core action. Right. Now I

404
00:19:58,400 --> 00:20:00,680
have this leftover prepositional phrase, the

405
00:20:00,680 --> 00:20:04,079
PP, with a telescope. Where does it go? In Tree

406
00:20:04,079 --> 00:20:07,140
1, I draw the branch for that PP high up, attaching

407
00:20:07,140 --> 00:20:10,079
it directly to the main verb phrase. It's modifying

408
00:20:10,079 --> 00:20:12,960
the verb saw. The geometry physically shows that

409
00:20:12,960 --> 00:20:14,940
the telescope is the instrument of the seeing.

410
00:20:15,460 --> 00:20:20,200
I used it to see him. Excellent. In Tree 2, I

411
00:20:20,200 --> 00:20:22,140
attach the branch for the prepositional phrase

412
00:20:22,140 --> 00:20:24,559
much lower down. I attach it directly inside

413
00:20:24,559 --> 00:20:27,960
the noun phrase for the astronomer. Lower attachment.

414
00:20:28,500 --> 00:20:31,349
Yeah. Now it's modifying the noun. The geometry

415
00:20:31,349 --> 00:20:33,549
shows that the telescope is a property belonging

416
00:20:33,549 --> 00:20:36,609
to the astronomer. The exact same string of words.

417
00:20:37,269 --> 00:20:40,410
Exactly. But two completely different, invisible

418
00:20:40,410 --> 00:20:42,349
geometries yielding two completely different

419
00:20:42,349 --> 00:20:46,230
meanings. It is incredibly elegant. CFG provides

420
00:20:46,230 --> 00:20:49,670
a beautiful visual logic for syntax. It really

421
00:20:49,670 --> 00:20:52,089
does. I was feeling pretty good about CFG when

422
00:20:52,089 --> 00:20:56,170
I read this part. But the textbook authors are

423
00:20:56,170 --> 00:20:59,289
ruthless about this. After spending pages teaching

424
00:20:59,289 --> 00:21:02,920
you CFG, they immediately tear it down. Wait,

425
00:21:02,920 --> 00:21:05,799
really? After all that beautiful geometry, what's

426
00:21:05,799 --> 00:21:09,099
wrong with it? They pivot and say, CFG is a great

427
00:21:09,099 --> 00:21:12,680
stepping stone, but it has a fatal flaw. A fatal

428
00:21:12,680 --> 00:21:16,579
flaw? CFG is actually too simple. It is far too

429
00:21:16,579 --> 00:21:20,619
permissive. How so? Well, in a standard context

430
00:21:20,619 --> 00:21:23,640
-free grammar, the categories like verb phrase

431
00:21:23,640 --> 00:21:26,940
or noun phrase are just opaque atomic symbols.

432
00:21:27,599 --> 00:21:30,160
What do you mean by atomic? The computer running

433
00:21:30,160 --> 00:21:33,880
a CFG just sees VP as a tag. It doesn't know

434
00:21:33,880 --> 00:21:37,920
what a verb actually is. Oh. Because of this,

435
00:21:38,220 --> 00:21:40,519
you could theoretically write a rule in CFG that

436
00:21:40,519 --> 00:21:43,680
says a verb phrase is composed of a preposition

437
00:21:43,680 --> 00:21:46,980
and a noun phrase. But that's absurd. A verb

438
00:21:46,980 --> 00:21:49,420
phrase has to have a verb in it. It's literally

439
00:21:49,420 --> 00:21:53,160
in the name. To a human, yes. It's obvious. But

440
00:21:53,160 --> 00:21:55,980
a pure CFG system has no mathematical mechanism

441
00:21:55,980 --> 00:21:59,150
to enforce that. Wow. It misses the fundamental

442
00:21:59,150 --> 00:22:02,470
cognitive concept of headedness. Headedness?

443
00:22:03,190 --> 00:22:05,750
In human language, every phrase is headed by

444
00:22:05,750 --> 00:22:09,289
its most important core word. A noun phrase centers

445
00:22:09,289 --> 00:22:11,869
around a noun. A prepositional phrase centers

446
00:22:11,869 --> 00:22:14,609
around a preposition. That makes perfect sense.

447
00:22:15,109 --> 00:22:17,970
But CFG forces linguists to write hundreds of

448
00:22:17,970 --> 00:22:21,190
redundant, overly specific rules for every possible

449
00:22:21,190 --> 00:22:24,250
combination because it can't grasp the simple

450
00:22:24,250 --> 00:22:26,869
generalization that Phrases inherit their properties

451
00:22:26,869 --> 00:22:30,509
from their core words. Okay. So we have to crack

452
00:22:30,509 --> 00:22:33,349
open these atomic symbols. We need to look inside

453
00:22:33,349 --> 00:22:36,470
the words themselves. Exactly. And that introduces

454
00:22:36,470 --> 00:22:39,950
a total paradigm shift in the textbook. Feature

455
00:22:39,950 --> 00:22:42,589
structures. Feature structures. We are no longer

456
00:22:42,589 --> 00:22:45,049
treating words as just solid blocks of text.

457
00:22:45,210 --> 00:22:48,089
We are treating them as bundles of traits. This

458
00:22:48,089 --> 00:22:51,829
is a profound upgrade. Instead of the word denies,

459
00:22:52,009 --> 00:22:54,430
just being tagged with a solitary V for verb,

460
00:22:54,890 --> 00:22:58,109
it becomes a complex matrix of features and values.

461
00:22:58,369 --> 00:23:01,170
Like a spreadsheet for a single word. Yes. It

462
00:23:01,170 --> 00:23:04,049
is like unpacking the genetic sequence, the DNA

463
00:23:04,049 --> 00:23:07,250
of the word. I really love this analogy. Let's

464
00:23:07,250 --> 00:23:09,690
look at the actual matrix. A feature structure

465
00:23:09,690 --> 00:23:12,390
is basically a mathematical function that maps

466
00:23:12,390 --> 00:23:15,450
a feature name to a specific value, right? That's

467
00:23:15,450 --> 00:23:17,710
the technical definition, yes. So if we look

468
00:23:17,710 --> 00:23:21,430
at the DNA of the verb denies, what do we actually

469
00:23:21,430 --> 00:23:24,630
see? Inside the matrix for denies, you will find

470
00:23:24,630 --> 00:23:27,569
a feature for tense, and its value is set to

471
00:23:27,569 --> 00:23:29,849
present. Okay. You'll find an agreement feature

472
00:23:29,849 --> 00:23:32,789
for its subject, requiring a third -person singular

473
00:23:32,789 --> 00:23:36,319
entity, like he or she. Right. And crucially,

474
00:23:36,700 --> 00:23:38,819
you will find a feature for its object, demanding

475
00:23:38,819 --> 00:23:42,079
that it must take a direct object. He denies

476
00:23:42,079 --> 00:23:44,660
the claim. Now, compare that to the DNA of the

477
00:23:44,660 --> 00:23:48,220
verb disappears. Good comparison. Both denies

478
00:23:48,220 --> 00:23:51,140
and disappears are verbs. Both take a third person

479
00:23:51,140 --> 00:23:54,259
singular subject. He denies works. He disappears

480
00:23:54,259 --> 00:23:57,859
works. Exactly. But inside the matrix for disappears,

481
00:23:58,059 --> 00:24:01,000
its object feature is set to none. Precisely.

482
00:24:01,480 --> 00:24:04,259
You could say he denies the claim, but you absolutely

483
00:24:04,259 --> 00:24:07,589
cannot say he disappears the claim. Unless you're

484
00:24:07,589 --> 00:24:10,690
in the mafia, I guess. Well, yes, apart from

485
00:24:10,690 --> 00:24:13,650
colloquial slang. But structurally, by treating

486
00:24:13,650 --> 00:24:16,730
words as feature matrices, we can cross -reference

487
00:24:16,730 --> 00:24:19,829
these traits. We can write broader rules. Exactly.

488
00:24:20,250 --> 00:24:22,809
We can write rules that say these two verbs share

489
00:24:22,809 --> 00:24:25,210
the subject feature but diverge on the object

490
00:24:25,210 --> 00:24:28,769
feature. We no longer need infinite redundant

491
00:24:28,769 --> 00:24:32,369
CFG rules. The words themselves contain the constraints.

492
00:24:32,759 --> 00:24:35,380
And this brings us to the master rule that holds

493
00:24:35,380 --> 00:24:38,660
this entire new framework together, the head

494
00:24:38,660 --> 00:24:42,259
feature principle, or HFP. The HFP is the bedrock

495
00:24:42,259 --> 00:24:45,019
of modern syntactic theory. Let's lay it out.

496
00:24:45,420 --> 00:24:47,859
It formally states that a phrase inherits the

497
00:24:47,859 --> 00:24:50,200
essential syntactic traits of its head daughter.

498
00:24:50,880 --> 00:24:54,099
Its head daughter. Meaning the core word. Right.

499
00:24:54,440 --> 00:24:57,180
The genetic material of the core word bubbles

500
00:24:57,180 --> 00:24:59,599
up through the tree to define the phrase above

501
00:24:59,599 --> 00:25:03,160
it. Give me an example. If the core noun at the

502
00:25:03,160 --> 00:25:05,880
bottom of the tree is plural, that plural feature

503
00:25:05,880 --> 00:25:08,539
bubbles up and the entire noun phrase becomes

504
00:25:08,539 --> 00:25:11,160
marked as plural in the overarching architecture.

505
00:25:12,160 --> 00:25:14,559
So this totally changes how we view a sentence

506
00:25:14,559 --> 00:25:17,420
tree. It's not just a static scaffolding we build

507
00:25:17,420 --> 00:25:19,700
over the words anymore. Not at all. The tree

508
00:25:19,700 --> 00:25:23,039
is alive. It's a constraint satisfaction network.

509
00:25:23,480 --> 00:25:25,619
Information is constantly flowing from the bottom

510
00:25:25,619 --> 00:25:28,549
up. A dynamic network. The word at the bottom

511
00:25:28,549 --> 00:25:31,829
passes its DNA up to the phrase level, enforcing

512
00:25:31,829 --> 00:25:34,210
mathematical rules and how that phrase can interact

513
00:25:34,210 --> 00:25:36,549
with the rest of the sentence. Which leads us

514
00:25:36,549 --> 00:25:39,470
perfectly to the next massive question. Yeah?

515
00:25:39,650 --> 00:25:42,049
If we have these feature structures, these bundles

516
00:25:42,049 --> 00:25:45,069
of DNA, how do they actually bind together to

517
00:25:45,069 --> 00:25:47,490
build a sentence? Right. How do the branches

518
00:25:47,490 --> 00:25:50,259
actually connect to each other? The authors introduce

519
00:25:50,259 --> 00:25:53,180
a feature called val, or valence. Valence, yes.

520
00:25:53,420 --> 00:25:55,660
Now, my mind immediately goes to high school

521
00:25:55,660 --> 00:25:58,900
chemistry when I hear that word. Valence is the

522
00:25:58,900 --> 00:26:01,200
combining capacity of an element, right? Exactly.

523
00:26:01,599 --> 00:26:03,980
An oxygen atom has two empty slots in its outer

524
00:26:03,980 --> 00:26:06,740
shell, so it chemically binds with two hydrogen

525
00:26:06,740 --> 00:26:09,799
atoms to become stable water. The chemical analogy

526
00:26:09,799 --> 00:26:12,400
is flawless, and it's exactly how the textbook

527
00:26:12,400 --> 00:26:15,940
approaches syntax. Words have a syntactic valence.

528
00:26:16,200 --> 00:26:18,900
They have empty slots. They have empty theoretical

529
00:26:18,900 --> 00:26:21,400
slots in their feature matrix that they absolutely

530
00:26:21,400 --> 00:26:24,539
must fill to become a complete, stable grammatical

531
00:26:24,539 --> 00:26:27,900
phrase. And how do we classify those slots? The

532
00:26:27,900 --> 00:26:29,940
valence feature groups together two main lists.

533
00:26:30,579 --> 00:26:33,299
SPR, which stands for specifiers. Things like

534
00:26:33,299 --> 00:26:36,420
subjects or determiners. Okay, SPR for specifiers.

535
00:26:36,920 --> 00:26:39,559
And COMPs, which stands for complements. Things

536
00:26:39,559 --> 00:26:42,900
like direct objects. Okay, so let's take a really

537
00:26:42,900 --> 00:26:47,289
hungry verb. Devour. Devour. Good one. If we

538
00:26:47,289 --> 00:26:50,029
look at the valence list in the DNA of devour,

539
00:26:50,529 --> 00:26:53,430
what does it need to become stable? Devour is

540
00:26:53,430 --> 00:26:56,170
highly unstable on its own. Its feature matrix

541
00:26:56,170 --> 00:26:59,529
says it needs a comps, a direct object, to consume.

542
00:26:59,789 --> 00:27:02,730
It needs something to be devoured. Right. It

543
00:27:02,730 --> 00:27:07,230
also needs an SPR, a subject, the devourer. It

544
00:27:07,230 --> 00:27:09,609
cannot exist in a sentence without both. Let's

545
00:27:09,609 --> 00:27:12,069
build a sentence step by step then. Let's say

546
00:27:12,069 --> 00:27:14,869
we have the words, the dog devoured the cake.

547
00:27:15,329 --> 00:27:19,049
How do the chemical bonds actually form? It happens

548
00:27:19,049 --> 00:27:21,809
systematically. The valence principle dictates

549
00:27:21,809 --> 00:27:25,029
the flow. Okay. First, the verb looks to its

550
00:27:25,029 --> 00:27:28,089
right. It finds its complement, the cake. It

551
00:27:28,089 --> 00:27:30,710
chemically binds with it, satisfying the comp's

552
00:27:30,710 --> 00:27:34,690
requirement. So, devour plus the cake. Once it

553
00:27:34,690 --> 00:27:37,049
consumes that complement, the verb graduates.

554
00:27:37,410 --> 00:27:41,109
It levels up from a simple verb to a verb phrase.

555
00:27:41,710 --> 00:27:44,779
Its comp's list is now empty and satisfied. But

556
00:27:44,779 --> 00:27:47,019
the whole phrase isn't stable yet because the

557
00:27:47,019 --> 00:27:49,880
SPR list, the subject slot, is still totally

558
00:27:49,880 --> 00:27:53,660
empty. Exactly. So the newly formed verb phrase

559
00:27:53,660 --> 00:27:56,960
looks to its left. It finds the specifier, the

560
00:27:56,960 --> 00:27:59,799
dog. And it binds with it. It binds with it.

561
00:28:00,180 --> 00:28:04,440
The dog devoured the cake. It consumes the specifier

562
00:28:04,440 --> 00:28:07,819
and now its SPR list is empty. And then what?

563
00:28:08,400 --> 00:28:10,660
When all the valence lists in the matrix are

564
00:28:10,660 --> 00:28:13,279
completely empty, the structure reaches total

565
00:28:13,279 --> 00:28:16,650
stability. It graduates to an S, a complete sentence.

566
00:28:17,289 --> 00:28:20,309
I am just picturing this cascade of chemical

567
00:28:20,309 --> 00:28:22,869
reactions firing off as a sentence is formed.

568
00:28:23,349 --> 00:28:25,490
But what happens if a constraint isn't met? What

569
00:28:25,490 --> 00:28:28,269
if one of the bonds fails? If a single feature

570
00:28:28,269 --> 00:28:31,069
clashes, say if the verb needs a plural subject,

571
00:28:31,309 --> 00:28:34,269
but the noun is singular, the constraint is not

572
00:28:34,269 --> 00:28:37,130
satisfied. The bond fails, the valence remains

573
00:28:37,130 --> 00:28:40,190
open, and the entire structure crashes. Total

574
00:28:40,190 --> 00:28:43,289
failure. The authors call this simultaneous satisfaction

575
00:28:43,289 --> 00:28:47,509
of all constraints, total well -typedness. Total

576
00:28:47,509 --> 00:28:50,900
well -typedness. Speaking a sentence isn't about

577
00:28:50,900 --> 00:28:53,319
running through a checklist of procedural rules.

578
00:28:53,720 --> 00:28:56,680
It is about instantly finding the one specific

579
00:28:56,680 --> 00:28:59,660
geometry where every single feature matrix aligns

580
00:28:59,660 --> 00:29:02,140
perfectly. That is genuinely awe -inspiring.

581
00:29:02,460 --> 00:29:05,500
It is. But it raises a huge logistical question

582
00:29:05,500 --> 00:29:07,960
for me. Where are all these features stored?

583
00:29:08,440 --> 00:29:10,849
What do you mean? If every single word has this

584
00:29:10,849 --> 00:29:14,130
incredibly complex matrix of tense, agreement,

585
00:29:14,410 --> 00:29:16,670
and valence, where does it live in the brain?

586
00:29:17,069 --> 00:29:19,730
It lives in the lexicon. The dictionary in our

587
00:29:19,730 --> 00:29:22,630
heads? Yes. And the textbook explains that the

588
00:29:22,630 --> 00:29:25,109
lexicon isn't just a flat alphabetical dictionary.

589
00:29:25,630 --> 00:29:28,509
It is a highly structured hierarchical database

590
00:29:28,509 --> 00:29:31,930
that utilizes multiple inheritance. We are moving

591
00:29:31,930 --> 00:29:34,970
beyond individual words and talking about lexemes.

592
00:29:35,309 --> 00:29:38,430
A lexeme. Okay, break that down for me. So, run...

593
00:29:38,569 --> 00:29:43,150
Runs, ran, running. These are all specific, realized

594
00:29:43,150 --> 00:29:46,309
word forms. The versions we actually speak. Right.

595
00:29:46,890 --> 00:29:49,210
But they all inherit their core DNA from the

596
00:29:49,210 --> 00:29:52,710
abstract root concept, the Lexi, for run. Got

597
00:29:52,710 --> 00:29:55,410
it. And by organizing the lexicon as a hierarchy,

598
00:29:55,789 --> 00:29:57,890
the architecture becomes incredibly efficient.

599
00:29:58,349 --> 00:30:01,529
Very efficient. Think of it like object -oriented

600
00:30:01,529 --> 00:30:03,750
programming in computer science. Oh, that's a

601
00:30:03,750 --> 00:30:06,269
perfect comparison. You can define a rule at

602
00:30:06,269 --> 00:30:09,119
the very top of the hierarchy. Say, at the level

603
00:30:09,119 --> 00:30:12,680
of all verbs. And every single specific verb

604
00:30:12,680 --> 00:30:15,279
lexeme underneath it automatically inherits that

605
00:30:15,279 --> 00:30:17,940
rule. You don't have to rewrite the rule thousands

606
00:30:17,940 --> 00:30:21,000
of times. Give me an example of a massive rule

607
00:30:21,000 --> 00:30:23,579
that sits in this hierarchy. The most important

608
00:30:23,579 --> 00:30:26,099
constraint operating in the lexicon is the shack.

609
00:30:26,980 --> 00:30:31,539
The shack. Sounds like a nightclub. The specifier

610
00:30:31,539 --> 00:30:35,200
head agreement constraint. OK. The shack. Let's

611
00:30:35,200 --> 00:30:37,599
look at the shack. This is the mechanism that

612
00:30:37,599 --> 00:30:40,799
ensures subjects and verbs match, right? That's

613
00:30:40,799 --> 00:30:44,700
the one. So the dog barks is a stable sentence,

614
00:30:44,720 --> 00:30:48,200
but the dog's barks crashes. How does the shack

615
00:30:48,200 --> 00:30:51,160
enforce this? The shack is a constraint written

616
00:30:51,160 --> 00:30:54,140
into the base definition of verbs. It forces

617
00:30:54,140 --> 00:30:57,180
the verb to mathematically examine the AGR, the

618
00:30:57,180 --> 00:30:59,359
agreement features, of whatever word is sitting

619
00:30:59,359 --> 00:31:02,859
in its specifier slot. So the verb is actively

620
00:31:02,859 --> 00:31:06,529
checking the subject. Yes. If the verb barks

621
00:31:06,529 --> 00:31:09,789
is marked as third -person singular, the SHAC

622
00:31:09,789 --> 00:31:12,589
reaches over and demands that its subject also

623
00:31:12,589 --> 00:31:15,109
be mathematically marked as third -person singular.

624
00:31:15,470 --> 00:31:19,089
And if it's not? If the subject is dogs, which

625
00:31:19,089 --> 00:31:22,390
is plural, the SHAC detects a clash. The valence

626
00:31:22,390 --> 00:31:25,650
combination is formally blocked. And the lexicon

627
00:31:25,650 --> 00:31:28,190
doesn't just store static words, does it? It

628
00:31:28,190 --> 00:31:30,970
also has rules for generating new words. Yes,

629
00:31:31,190 --> 00:31:34,220
morphological rules. The textbook draws a really

630
00:31:34,220 --> 00:31:36,559
sharp distinction between inflectional rules

631
00:31:36,559 --> 00:31:38,759
and derivational rules. I want to make sure I

632
00:31:38,759 --> 00:31:41,359
grasp the difference. It's an important distinction.

633
00:31:42,329 --> 00:31:45,349
Inflectional rules are minor tweaks. Minor tweaks.

634
00:31:45,609 --> 00:31:48,109
They change the form of a word, like its tense

635
00:31:48,109 --> 00:31:50,589
or its number, without changing its fundamental

636
00:31:50,589 --> 00:31:53,630
part of speech or its core meaning, turning the

637
00:31:53,630 --> 00:31:57,150
verb walk into walked or the noun cat into cats.

638
00:31:57,890 --> 00:32:00,529
The valence remains exactly the same. Right.

639
00:32:01,250 --> 00:32:04,049
Derivational rules, however, are massive structural

640
00:32:04,049 --> 00:32:06,750
mutations. Structural mutations. They change

641
00:32:06,750 --> 00:32:09,210
the fundamental category or the argument structure

642
00:32:09,210 --> 00:32:12,250
of the word itself. The text uses the passive

643
00:32:12,250 --> 00:32:14,849
transformation as the ultimate example of a derivational

644
00:32:14,849 --> 00:32:18,009
rule, right? The passive voice, yes. Taking an

645
00:32:18,009 --> 00:32:20,150
active sentence like the dog bit the man and

646
00:32:20,150 --> 00:32:22,130
turning it into the man was bitten by the dog.

647
00:32:22,490 --> 00:32:25,049
Exactly. I remember in older linguistic theories

648
00:32:25,049 --> 00:32:27,329
they used to draw these diagrams where the actual

649
00:32:27,329 --> 00:32:30,109
words physically moved across the sentence, like

650
00:32:30,109 --> 00:32:33,150
chess pieces. Yes, older transformational grammars

651
00:32:33,150 --> 00:32:35,769
relied heavily on invisible movement. Right.

652
00:32:36,069 --> 00:32:38,509
But the beauty of constraint -based lexicalism,

653
00:32:38,849 --> 00:32:41,470
the theory this textbook champions, is that there

654
00:32:41,470 --> 00:32:45,049
is no movement. No movement at all. None. The

655
00:32:45,049 --> 00:32:47,250
passive transformation is just a derivational

656
00:32:47,250 --> 00:32:50,269
rule that happens deep inside the lexicon, before

657
00:32:50,269 --> 00:32:53,049
the words ever enter a sentence tree. How does

658
00:32:53,049 --> 00:32:56,230
that work? How does the lexicon passively transform

659
00:32:56,230 --> 00:32:58,920
a word without moving it? The derivational rule

660
00:32:58,920 --> 00:33:02,420
takes the active verb lexeme, bite, and mathematically

661
00:33:02,420 --> 00:33:05,220
scrambles its valence list. It rearranges the

662
00:33:05,220 --> 00:33:09,039
DNA. Exactly. It spits out a brand new mutated

663
00:33:09,039 --> 00:33:13,160
passive verb lexeme, bitten. This new lexeme

664
00:33:13,160 --> 00:33:15,359
has completely different chemical properties.

665
00:33:15,799 --> 00:33:18,680
Its matrix now dictates that the old object must

666
00:33:18,680 --> 00:33:21,420
become its new subject, and the old subject gets

667
00:33:21,420 --> 00:33:23,539
pushed into an optional prepositional phrase.

668
00:33:24,019 --> 00:33:27,390
By the dog. Right. All the heavy lifting is done

669
00:33:27,390 --> 00:33:30,269
inside the word's DNA. The tree building itself

670
00:33:30,269 --> 00:33:33,329
remains pure, simple, and entirely surface -oriented.

671
00:33:33,869 --> 00:33:36,970
It's just breathtaking how elegant the math is.

672
00:33:37,210 --> 00:33:40,130
It's very satisfying. But I do have a nagging

673
00:33:40,130 --> 00:33:42,109
thought. We have spent all this time talking

674
00:33:42,109 --> 00:33:44,990
about syntax. Matrices, valence, constraints,

675
00:33:45,170 --> 00:33:48,289
trees. But syntax is just structure. That's true.

676
00:33:48,549 --> 00:33:50,769
A sentence is totally useless if it doesn't actually

677
00:33:50,769 --> 00:33:53,609
mean anything. How does meaning plug into this

678
00:33:53,609 --> 00:33:57,130
system? A vital question. Syntax without meaning

679
00:33:57,130 --> 00:34:00,289
is just sterile algebra. Right. The textbook

680
00:34:00,289 --> 00:34:02,670
tackles this head -on when exploring semantics.

681
00:34:03,329 --> 00:34:05,829
They demonstrate how meaning maps directly onto

682
00:34:05,829 --> 00:34:08,170
these feature structures. How do they break it

683
00:34:08,170 --> 00:34:11,510
down? They break human meaning down into three

684
00:34:11,510 --> 00:34:14,409
abstract semantic objects. Propositions, which

685
00:34:14,409 --> 00:34:16,769
are statements that can be true or false. Like,

686
00:34:16,869 --> 00:34:19,570
the sky is blue? Right. Questions, which can

687
00:34:19,570 --> 00:34:22,489
be asked and answered. And directives, which

688
00:34:22,489 --> 00:34:25,400
can be issued and fulfilled, like commands. And

689
00:34:25,400 --> 00:34:26,980
when I was looking at the diagrams in the book,

690
00:34:27,119 --> 00:34:29,739
I noticed that inside the massive feature matrix

691
00:34:29,739 --> 00:34:32,880
for a word, right next to the syntax block, there

692
00:34:32,880 --> 00:34:35,679
is a whole separate semantics block. Yes, they

693
00:34:35,679 --> 00:34:37,980
run in parallel. And it has these really specific

694
00:34:37,980 --> 00:34:42,460
lists. The SIT for situation, the index for specific

695
00:34:42,460 --> 00:34:45,880
entities, and the REST for restrictions. Exactly.

696
00:34:46,239 --> 00:34:48,820
Let's break those down. The index feature provides

697
00:34:48,820 --> 00:34:51,800
a unique mathematical ID number for the specific

698
00:34:51,800 --> 00:34:54,179
person or thing being talked about. Like a barcode.

699
00:34:54,340 --> 00:34:58,019
Like a barcode. The RESTR list is a set of conditions

700
00:34:58,019 --> 00:35:01,239
that apply to that index. I found a metaphor

701
00:35:01,239 --> 00:35:03,159
for this that really helped me visualize it.

702
00:35:03,340 --> 00:35:06,019
Let's hear it. I like to think of this as casting

703
00:35:06,019 --> 00:35:09,539
a play. The syntax block sets the stage. It determines

704
00:35:09,539 --> 00:35:12,179
who stands where, who is the subject, who is

705
00:35:12,179 --> 00:35:15,820
the object. But the semantic RESTR list is the

706
00:35:15,820 --> 00:35:18,480
script. It assigns the actual roles. That's a

707
00:35:18,480 --> 00:35:21,019
fantastic conceptualization. It helps it click

708
00:35:21,019 --> 00:35:24,730
for me. If we take the sentence, Kim likes Sandy.

709
00:35:25,610 --> 00:35:28,210
The syntax builds the tree and satisfies the

710
00:35:28,210 --> 00:35:32,969
valence. But simultaneously, the RESTR list steps

711
00:35:32,969 --> 00:35:37,590
in. The script. It says, okay, we have a liking

712
00:35:37,590 --> 00:35:40,730
situation here. I am formally assigning the role

713
00:35:40,730 --> 00:35:44,389
of liker to the index belonging to Kim, and I'm

714
00:35:44,389 --> 00:35:46,989
assigning the role of liked to the index belonging

715
00:35:46,989 --> 00:35:50,570
to Sandy. It binds the abstract syntactic slots

716
00:35:50,570 --> 00:35:53,869
to the real -world actors. Exactly. And once

717
00:35:53,869 --> 00:35:56,750
we have this indexing power, this ability to

718
00:35:56,750 --> 00:35:59,949
assign numerical IDs to actors, the textbook

719
00:35:59,949 --> 00:36:02,809
does something incredibly satisfying. Oh, I know

720
00:36:02,809 --> 00:36:04,969
what you're going to say. It circles all the

721
00:36:04,969 --> 00:36:07,489
way back to the reflexive pronoun puzzle we started

722
00:36:07,489 --> 00:36:10,309
the deep dive with. Binding theory. Binding theory.

723
00:36:11,340 --> 00:36:13,780
Earlier, we informally solved the puzzle with

724
00:36:13,780 --> 00:36:16,860
hypothesis three, saying reflexive pronouns must

725
00:36:16,860 --> 00:36:19,579
be co -arguments of the verb. Right, the whole

726
00:36:19,579 --> 00:36:22,400
core participants thing. But now, armed with

727
00:36:22,400 --> 00:36:25,519
feature matrices and semantic indexing, the authors

728
00:36:25,519 --> 00:36:28,719
can formally mathematically prove it. How do

729
00:36:28,719 --> 00:36:31,119
they do it? Walk me through the math. They introduce

730
00:36:31,119 --> 00:36:33,840
a new list in the feature matrix called the argument

731
00:36:33,840 --> 00:36:38,480
structure, or ARGST. This list ranks the arguments

732
00:36:38,480 --> 00:36:41,559
of a verb by prominence, subject first, object

733
00:36:41,559 --> 00:36:44,800
second, and so on. The formal rule of binding

734
00:36:44,800 --> 00:36:47,139
theory states that the grammar actually forces

735
00:36:47,139 --> 00:36:50,719
a reflexive pronoun to share an index, a semantic

736
00:36:50,719 --> 00:36:53,780
ID number, with a preceding, more prominent argument

737
00:36:53,780 --> 00:36:56,639
on the exact same argist he lists. It forces

738
00:36:56,639 --> 00:36:59,260
it. If the pronoun's matrix searches that specific

739
00:36:59,260 --> 00:37:01,679
list and doesn't find a matching index to bind

740
00:37:01,679 --> 00:37:04,260
to, the constraint fails and the sentence crashes.

741
00:37:04,679 --> 00:37:07,179
It's so deeply satisfying when the rigorous math

742
00:37:07,179 --> 00:37:09,659
perfectly models that initial human intuition

743
00:37:09,659 --> 00:37:12,119
we had. It's the moment it all comes together.

744
00:37:12,219 --> 00:37:14,539
And from there, the textbook really takes the

745
00:37:14,539 --> 00:37:16,579
training wheels off. The next several chapters,

746
00:37:16,619 --> 00:37:19,380
I think it's 10 through 15, are essentially a

747
00:37:19,380 --> 00:37:22,199
gauntlet. A gauntlet is the right word. The authors

748
00:37:22,199 --> 00:37:24,480
just throw the weirdest, most complex anomalies

749
00:37:24,480 --> 00:37:26,239
of the English language at the theory to see

750
00:37:26,239 --> 00:37:28,300
if the architecture holds up. They are stress

751
00:37:28,300 --> 00:37:31,519
testing the model. And one of the most fascinating

752
00:37:31,519 --> 00:37:35,000
edge cases they examine is the dummy there. The

753
00:37:35,000 --> 00:37:37,860
dummy there. As in the sentence, there is a seat

754
00:37:37,860 --> 00:37:41,059
available. Exactly. I... have literally never

755
00:37:41,059 --> 00:37:43,679
thought about how weird the word there is in

756
00:37:43,679 --> 00:37:46,219
that context. Because it doesn't mean a physical

757
00:37:46,219 --> 00:37:48,559
location. You aren't pointing across the room

758
00:37:48,559 --> 00:37:51,800
and saying, over there is a seat. You're just

759
00:37:51,800 --> 00:37:55,739
stating the existence of the seat. Exactly. In

760
00:37:55,739 --> 00:37:58,300
terms of our semantic matrix, the dummy there

761
00:37:58,300 --> 00:38:02,000
is a ghost. A ghost? It lacks referential meaning

762
00:38:02,000 --> 00:38:05,679
entirely. It has no index. Wait, really? No ID

763
00:38:05,679 --> 00:38:09,139
number? None. But syntactically, the verb is?

764
00:38:09,260 --> 00:38:11,800
demands its subject to satisfy its valence constraints.

765
00:38:12,300 --> 00:38:15,400
It absolutely needs its SPR slot filled or the

766
00:38:15,400 --> 00:38:19,159
sentence crashes. So the word there just steps

767
00:38:19,159 --> 00:38:22,519
in as a pure structural placeholder? Yes. It's

768
00:38:22,519 --> 00:38:25,440
a syntactic dummy. It fills the valence slot,

769
00:38:25,780 --> 00:38:29,380
it satisfies the SHAC agreement, but it contributes

770
00:38:29,380 --> 00:38:32,159
absolutely nothing to the semantic RESTR list?

771
00:38:32,460 --> 00:38:35,880
Correct. The system is robust enough to process

772
00:38:35,880 --> 00:38:37,860
words that are essentially just mathematical

773
00:38:37,860 --> 00:38:41,349
filler. That's so cool. What else do they test?

774
00:38:41,590 --> 00:38:43,889
They apply that same rigorous logic to other

775
00:38:43,889 --> 00:38:47,110
anomalies. They prove that the infinitival 2,

776
00:38:47,389 --> 00:38:50,949
like in I want to go, isn't just a random particle.

777
00:38:51,150 --> 00:38:53,670
Right. Its feature structure proves it acts as

778
00:38:53,670 --> 00:38:56,860
an auxiliary verb. But the ultimate stress test

779
00:38:56,860 --> 00:38:59,280
they perform is on long -distance dependencies,

780
00:38:59,619 --> 00:39:02,579
or LDDs. Okay, I remember staring at the pages

781
00:39:02,579 --> 00:39:05,000
on LDDs and literally feeling my brain start

782
00:39:05,000 --> 00:39:07,329
to sweat. It's dense. This is one of where it

783
00:39:07,329 --> 00:39:09,750
gets moved far away from its natural grammatical

784
00:39:09,750 --> 00:39:12,889
home, right? Like in a question. Correct. Consider

785
00:39:12,889 --> 00:39:16,769
the sentence, Who did you say that Kim saw? Who

786
00:39:16,769 --> 00:39:19,590
did you say that Kim saw? Think about the verb

787
00:39:19,590 --> 00:39:23,269
saw at the very end. Saw is a transitive verb.

788
00:39:23,469 --> 00:39:26,369
Its valence demands a direct object. But there

789
00:39:26,369 --> 00:39:28,670
is no object after it. The sentence just ends.

790
00:39:29,150 --> 00:39:32,369
Exactly. The object is who, and it's sitting

791
00:39:32,369 --> 00:39:34,840
all the way at the front of the sentence. So

792
00:39:34,840 --> 00:39:37,139
how does the tree not crash? The valence of saw

793
00:39:37,139 --> 00:39:40,079
is totally unsatisfied. It's hungry for a compliment.

794
00:39:40,659 --> 00:39:43,260
To solve this, the authors introduce a brilliant

795
00:39:43,260 --> 00:39:46,039
mathematical mechanism called the gap feature.

796
00:39:46,300 --> 00:39:49,619
The gap feature. When the verb saw realizes its

797
00:39:49,619 --> 00:39:52,420
object is missing, it doesn't crash. Instead,

798
00:39:52,840 --> 00:39:55,400
it generates a gap feature in its matrix. It

799
00:39:55,400 --> 00:39:58,769
leaves a mathematical memory trail. So saw passes

800
00:39:58,769 --> 00:40:01,889
this gap feature up to its mother node, the verb

801
00:40:01,889 --> 00:40:04,449
phrase. Yes. And then the verb phrase passes

802
00:40:04,449 --> 00:40:07,449
it up to the sentence node. I can just trace

803
00:40:07,449 --> 00:40:09,329
the signal climbing up the branches of the tree.

804
00:40:09,929 --> 00:40:12,750
Precisely. The gap feature bubbles up node by

805
00:40:12,750 --> 00:40:15,469
node, carrying the exact specifications of the

806
00:40:15,469 --> 00:40:18,210
missing object. It keeps climbing until it reaches

807
00:40:18,210 --> 00:40:20,829
the very top of the tree, where it finally encounters

808
00:40:20,829 --> 00:40:24,309
the word who. And who has exactly the features

809
00:40:24,309 --> 00:40:27,500
the gap is looking for. Yes. And at that exact

810
00:40:27,500 --> 00:40:30,559
node, the grammar rules introduce a stopgap feature.

811
00:40:31,159 --> 00:40:34,400
Stopgap. The stopgap tells the tree, the missing

812
00:40:34,400 --> 00:40:37,260
piece has been found. The dependency is resolved.

813
00:40:37,619 --> 00:40:40,840
You can stop searching now. The gap is mathematically

814
00:40:40,840 --> 00:40:43,800
closed, the valence is retrospectively satisfied,

815
00:40:44,280 --> 00:40:47,550
and the sentence is grammatical. It is a stunning

816
00:40:47,550 --> 00:40:49,789
piece of theoretical engineering. It really is

817
00:40:49,789 --> 00:40:51,869
beautiful. And what's really important to highlight

818
00:40:51,869 --> 00:40:53,869
here, especially for the listener, is that the

819
00:40:53,869 --> 00:40:55,909
textbook doesn't just use this engineering to

820
00:40:55,909 --> 00:40:58,630
analyze standard formal English. No, it doesn't.

821
00:40:58,929 --> 00:41:02,289
It applies this exact same rigorous math to intra

822
00:41:02,289 --> 00:41:04,349
-language variations. Specifically, they do a

823
00:41:04,349 --> 00:41:06,630
deep dive into African American Vernacular English,

824
00:41:06,849 --> 00:41:09,590
or AAVE. Yes, and this is a crucial demonstration

825
00:41:09,590 --> 00:41:12,010
of the core premise we established in the beginning.

826
00:41:12,769 --> 00:41:15,750
Linguistics is descriptive. not prescriptive.

827
00:41:16,110 --> 00:41:18,090
Right. We aren't judging. We're observing the

828
00:41:18,090 --> 00:41:20,550
anatomy. It observes natural language systems

829
00:41:20,550 --> 00:41:24,809
without judgment. And AAVE possesses a very specific,

830
00:41:24,849 --> 00:41:27,550
well -documented grammatical feature known as

831
00:41:27,550 --> 00:41:30,369
copula absence. Right. So in standard English,

832
00:41:30,449 --> 00:41:33,769
a speaker might say, it is wild or she is happy.

833
00:41:34,170 --> 00:41:37,030
In AAVE, a native speaker might say, it wild

834
00:41:37,030 --> 00:41:40,170
or she happy. Historically, prescriptivists,

835
00:41:40,449 --> 00:41:43,300
the dress code enforcers, tried to label this

836
00:41:43,300 --> 00:41:46,260
as broken English. Which is just so wrong. They

837
00:41:46,260 --> 00:41:48,619
claim the speaker simply forgot the verb or was

838
00:41:48,619 --> 00:41:51,780
speaking lazily. But the textbook applies the

839
00:41:51,780 --> 00:41:54,179
exact same constraint -based feature architecture

840
00:41:54,179 --> 00:41:57,300
to AAVE that it applies to the Queen's English.

841
00:41:57,559 --> 00:41:59,840
And the mathematical results are undeniable.

842
00:42:00,719 --> 00:42:03,019
What does the feature matrix reveal? It reveals

843
00:42:03,019 --> 00:42:06,519
that AAVE is governed by incredibly strict, systematic,

844
00:42:06,699 --> 00:42:10,280
and complex rules. Copula absence isn't a missing

845
00:42:10,280 --> 00:42:13,289
constituent at all. The grammar of AAVE formally

846
00:42:13,289 --> 00:42:16,909
licenses what linguists call a silent verb. A

847
00:42:16,909 --> 00:42:19,710
silent verb. Yeah. That sounds like an oxymoron,

848
00:42:19,730 --> 00:42:22,090
but in this architecture, it makes perfect sense.

849
00:42:22,570 --> 00:42:24,670
It's a word with a full feature structure. It

850
00:42:24,670 --> 00:42:27,329
has valence constraints. It has semantic indexing.

851
00:42:27,610 --> 00:42:29,909
It just has a blank phonological output. It makes

852
00:42:29,909 --> 00:42:33,489
no sound. Exactly. The silent verb in AAVE still

853
00:42:33,489 --> 00:42:36,309
demands a specific subject to fill its SPR slot.

854
00:42:36,769 --> 00:42:38,530
It still demands a complement to fill its comp

855
00:42:38,530 --> 00:42:42,000
slot. it still strictly enforces agreement constraints.

856
00:42:42,179 --> 00:42:45,039
Right. If you break those constraints, an AAVE

857
00:42:45,039 --> 00:42:47,039
speaker will instantly recognize the sentence

858
00:42:47,039 --> 00:42:49,639
as ungrammatical within their dialect. Because

859
00:42:49,639 --> 00:42:52,260
it violates the internal math. The theoretical

860
00:42:52,260 --> 00:42:54,719
architecture perfectly maps the hidden, rigorous

861
00:42:54,719 --> 00:42:57,039
rules of diverse dialects without an ounce of

862
00:42:57,039 --> 00:42:59,619
bias. It's the ultimate proof that the anatomy

863
00:42:59,619 --> 00:43:01,500
is real, regardless of what clothes you wear

864
00:43:01,500 --> 00:43:04,360
over it. It's just beautiful science. It is.

865
00:43:04,590 --> 00:43:08,010
Okay, we have built this massive, towering mathematical

866
00:43:08,010 --> 00:43:10,510
model. But the textbook doesn't stop there. In

867
00:43:10,510 --> 00:43:12,650
Chapter 9, they take all this beautiful abstract

868
00:43:12,650 --> 00:43:15,329
algebra and they slam it directly into biological

869
00:43:15,329 --> 00:43:18,889
reality. Psycholinguistics. Right. Psycholinguistics.

870
00:43:19,369 --> 00:43:21,489
The authors argue that this specific theory,

871
00:43:21,909 --> 00:43:24,389
constraint -based lexicalism, or CBL, isn't just

872
00:43:24,389 --> 00:43:27,550
a neat mathematical trick. They argue it is psychologically

873
00:43:27,550 --> 00:43:30,849
real. It is the actual literal software the human

874
00:43:30,849 --> 00:43:33,530
brain runs. To understand why this is such a

875
00:43:33,530 --> 00:43:36,610
bold claim, we have to revisit Chomsky's famous

876
00:43:36,610 --> 00:43:39,590
distinction between competence and performance.

877
00:43:40,510 --> 00:43:42,889
Competence is your abstract underlying knowledge

878
00:43:42,889 --> 00:43:45,670
of the language. It's the perfect, pristine feature

879
00:43:45,670 --> 00:43:47,510
matrices we've been building for the last hour.

880
00:43:47,809 --> 00:43:50,409
Yes. And performance is what your brain actually

881
00:43:50,409 --> 00:43:53,510
does in real time when you are trying to understand

882
00:43:53,510 --> 00:43:55,530
a fast -talking friend in a noisy restaurant.

883
00:43:55,969 --> 00:43:59,389
Performance is messy. Extremely messy. It's constrained

884
00:43:59,389 --> 00:44:02,469
by memory limits and processing speed. The holy

885
00:44:02,469 --> 00:44:04,909
grail of cognitive linguistics is to build a

886
00:44:04,909 --> 00:44:07,650
competence model that seamlessly translates into

887
00:44:07,650 --> 00:44:10,250
a performance model. A bridge between theory

888
00:44:10,250 --> 00:44:12,590
and reality. And the authors use garden path

889
00:44:12,590 --> 00:44:15,610
sentences to prove that CBL achieves this. Garden

890
00:44:15,610 --> 00:44:18,030
path sentences are my absolute favorite. They're

891
00:44:18,030 --> 00:44:20,670
very revealing. They literally make you feel

892
00:44:20,670 --> 00:44:23,050
your brain crashing in real time. The classic

893
00:44:23,050 --> 00:44:25,750
example in the text is, the horse raced past

894
00:44:25,750 --> 00:44:28,849
the barn fell. Let's slow down and trace exactly

895
00:44:28,849 --> 00:44:31,349
what happens in the listener's brain as they

896
00:44:31,349 --> 00:44:34,349
process that sentence word by word. Okay. You

897
00:44:34,349 --> 00:44:37,230
hear the horse. Your brain instantly opens a

898
00:44:37,230 --> 00:44:40,469
noun phrase structure. You hear raced. Your brain

899
00:44:40,469 --> 00:44:43,090
analyzes the statistical probability of the word

900
00:44:43,090 --> 00:44:45,889
raced, decides it is almost always an active

901
00:44:45,889 --> 00:44:48,150
verb, and slots it into the tree as the main

902
00:44:48,150 --> 00:44:51,409
action. The horse is racing. You hear past the

903
00:44:51,409 --> 00:44:54,150
barn. Your brain attaches it as a prepositional

904
00:44:54,150 --> 00:44:56,679
phrase, modifying the racing. Everything is going

905
00:44:56,679 --> 00:44:59,360
great. The tree is building perfectly. I can

906
00:44:59,360 --> 00:45:01,360
see the horse running past the barn. But then

907
00:45:01,360 --> 00:45:04,920
you hit the last word, fell. And the entire architecture

908
00:45:04,920 --> 00:45:08,579
collapses. The horse raced past the barn, fell.

909
00:45:09,380 --> 00:45:12,199
You have a spare, unattached verb with no subject.

910
00:45:12,840 --> 00:45:15,760
The valence constraints clash violently. You

911
00:45:15,760 --> 00:45:17,880
were led down a garden path. You thought you

912
00:45:17,880 --> 00:45:20,199
were in a simple act of sentence. But you have

913
00:45:20,199 --> 00:45:22,360
to hit the brakes, rewind the tape, and reparse

914
00:45:22,360 --> 00:45:24,880
the entire geometry of the sentence. Exactly.

915
00:45:25,150 --> 00:45:28,010
You realize, raced wasn't the main verb at all.

916
00:45:28,050 --> 00:45:30,690
It was a passive modifier. The true underlying

917
00:45:30,690 --> 00:45:33,789
structure is, the horse that was raced past the

918
00:45:33,789 --> 00:45:37,130
barn by someone fell. But why did you crash in

919
00:45:37,130 --> 00:45:39,809
the first place? Well, because we process language

920
00:45:39,809 --> 00:45:42,630
incrementally. Right. We don't wait for a person

921
00:45:42,630 --> 00:45:44,829
to finish their entire sentence, hit enter, and

922
00:45:44,829 --> 00:45:47,090
then compile the tree. We are building the tree

923
00:45:47,090 --> 00:45:50,050
in real time, word by word, guessing the geometry

924
00:45:50,050 --> 00:45:53,179
based on probabilities. Precisely. And the textbook

925
00:45:53,179 --> 00:45:55,679
cites fascinating eye -tracking experiments to

926
00:45:55,679 --> 00:45:57,860
prove how tightly integrated this process is.

927
00:45:58,019 --> 00:46:00,920
Tracking where people look? Yes. They track exactly

928
00:46:00,920 --> 00:46:03,539
where a person's eyes move as they read, pinpointing

929
00:46:03,539 --> 00:46:06,760
the exact millisecond they get confused. These

930
00:46:06,760 --> 00:46:09,500
experiments prove that humans use semantic plausibility

931
00:46:09,500 --> 00:46:12,639
to instantly resolve syntactic ambiguity on the

932
00:46:12,639 --> 00:46:15,199
fly. The comparison they use is brilliant. They

933
00:46:15,199 --> 00:46:18,179
contrast two sentences. Sentence A, the thieves

934
00:46:18,179 --> 00:46:21,239
assembled. Okay. And sentence B, the evidence

935
00:46:21,239 --> 00:46:23,969
assembled. Now, syntactically, those are identical.

936
00:46:24,230 --> 00:46:26,389
It's a plural noun followed by the word assembled.

937
00:46:26,630 --> 00:46:29,050
But semantically, they are vastly different.

938
00:46:29,530 --> 00:46:31,429
Thieves are animate human beings. They have the

939
00:46:31,429 --> 00:46:34,170
physical agency to gather together. Right. So

940
00:46:34,170 --> 00:46:37,150
when the brain hears the thieves assembled, it

941
00:46:37,150 --> 00:46:40,489
instantly builds an active verb tree. But evidence

942
00:46:40,489 --> 00:46:43,710
is inanimate. A pile of documents cannot physically

943
00:46:43,710 --> 00:46:46,179
assemble itself. It cannot. So when the brain

944
00:46:46,179 --> 00:46:49,440
processes the evidence assembled, it instantly

945
00:46:49,440 --> 00:46:52,719
detects a semantic violation. The RESTR list

946
00:46:52,719 --> 00:46:55,440
clashes with the real world. And here is the

947
00:46:55,440 --> 00:46:58,039
profound part. The eye tracking data shows that

948
00:46:58,039 --> 00:47:00,480
readers never get led down a garden path with

949
00:47:00,480 --> 00:47:03,000
the evidence assembled. They don't even blink.

950
00:47:03,480 --> 00:47:05,800
Because the semantic implausibility is detected

951
00:47:05,800 --> 00:47:08,659
instantly, the brain immediately abandons the

952
00:47:08,659 --> 00:47:11,159
active verb tree and automatically parses it

953
00:47:11,159 --> 00:47:13,909
as a passive modifier. the evidence that was

954
00:47:13,909 --> 00:47:17,409
assembled by the police. Exactly. That is staggering.

955
00:47:17,869 --> 00:47:20,510
It means that syntax and semantics aren't two

956
00:47:20,510 --> 00:47:22,510
different assembly lines in the brain. No, they

957
00:47:22,510 --> 00:47:24,550
aren't. You don't build the tree first and then

958
00:47:24,550 --> 00:47:27,090
check the meaning later. They are happening completely

959
00:47:27,090 --> 00:47:29,769
in parallel. And this is exactly why constraint

960
00:47:29,769 --> 00:47:32,530
-based lexicalism is biologically plausible.

961
00:47:33,170 --> 00:47:34,949
Because in this model, there are no step -by

962
00:47:34,949 --> 00:47:38,059
-step sequential rules. It's all at once. The

963
00:47:38,059 --> 00:47:40,739
syntax, the valence, and the semantics are all

964
00:47:40,739 --> 00:47:43,400
inextricably bundled together inside a word's

965
00:47:43,400 --> 00:47:46,849
DNA. As each word enters the brain, all of those

966
00:47:46,849 --> 00:47:49,869
constraints are evaluated simultaneously in a

967
00:47:49,869 --> 00:47:52,949
massive parallel network. The math of the textbook

968
00:47:52,949 --> 00:47:55,550
perfectly mimics the incremental integrative

969
00:47:55,550 --> 00:47:57,690
biology of the brain. It makes you look at a

970
00:47:57,690 --> 00:47:59,590
simple conversation in a completely different

971
00:47:59,590 --> 00:48:02,349
light, doesn't it? It really does. The sheer

972
00:48:02,349 --> 00:48:06,300
processing power is astronomical. Okay, we are

973
00:48:06,300 --> 00:48:09,920
entering the final chapter. Chapter 16, the ultimate

974
00:48:09,920 --> 00:48:12,960
synthesis. The grand finale. And I have to say,

975
00:48:13,219 --> 00:48:15,760
the authors pull off an absolute theoretical

976
00:48:15,760 --> 00:48:18,739
magic trick here. They do indeed. After spending

977
00:48:18,739 --> 00:48:21,559
15 densely packed chapters teaching the reader

978
00:48:21,559 --> 00:48:24,440
all these grammar rules, the head compliment

979
00:48:24,440 --> 00:48:27,380
rule, the head specifier rule, the rules that

980
00:48:27,380 --> 00:48:29,960
build the trees. Yeah. They suddenly discard

981
00:48:29,960 --> 00:48:32,739
the concept of procedural rules entirely. I was

982
00:48:32,739 --> 00:48:35,369
reading this thinking, wait, what? You're throwing

983
00:48:35,369 --> 00:48:38,250
out the rules? I just spent hours learning them.

984
00:48:38,650 --> 00:48:41,469
They don't throw away the logic. They unify it.

985
00:48:41,889 --> 00:48:44,610
They introduce sign -based construction grammar.

986
00:48:45,150 --> 00:48:48,150
Sign -based construction grammar. Instead of

987
00:48:48,150 --> 00:48:50,730
having a procedural rule that takes pieces and

988
00:48:50,730 --> 00:48:53,429
builds a tree, they introduce constructions.

989
00:48:53,909 --> 00:48:56,929
A construction is just a massive, highly abstract

990
00:48:56,929 --> 00:49:00,050
feature matrix. Okay. And inside this matrix,

991
00:49:00,210 --> 00:49:02,969
there are two new lists. A mother feature and

992
00:49:02,969 --> 00:49:05,230
a daughter's feature. Okay, let me make sure

993
00:49:05,230 --> 00:49:07,909
I'm visualizing this correctly. Up until now,

994
00:49:08,110 --> 00:49:10,570
a single word was a feature matrix. But in this

995
00:49:10,570 --> 00:49:13,829
final synthesis, an entire phrase, even an entire

996
00:49:13,829 --> 00:49:17,090
sentence, is also just treated as one massive

997
00:49:17,090 --> 00:49:20,829
feature matrix. Yes. They eliminate the distinction

998
00:49:20,829 --> 00:49:23,690
between words and phrases. Everything is just

999
00:49:23,690 --> 00:49:27,050
a sign. A sign. A sentence is no longer the result

1000
00:49:27,050 --> 00:49:30,190
of a step -by -step building process. The grammar

1001
00:49:30,190 --> 00:49:32,650
is no longer an active fact reassembly line.

1002
00:49:32,829 --> 00:49:35,929
It is a passive filter. Walk me through the filter

1003
00:49:35,929 --> 00:49:38,909
analogy. How does a construction act as a filter?

1004
00:49:39,570 --> 00:49:42,090
Imagine you take an utterance, any random string

1005
00:49:42,090 --> 00:49:44,429
of words. You drop it through the construction

1006
00:49:44,429 --> 00:49:47,550
filter. The filter doesn't build anything. It

1007
00:49:47,550 --> 00:49:50,510
just performs a simultaneous mathematical check.

1008
00:49:51,329 --> 00:49:54,050
Checking what against what? Do the aggregate

1009
00:49:54,050 --> 00:49:56,969
features of the daughters seamlessly satisfy

1010
00:49:56,969 --> 00:49:59,030
all the structural constraints demanded by the

1011
00:49:59,030 --> 00:50:02,150
mother? Does the valence perfectly lock together?

1012
00:50:02,630 --> 00:50:04,909
Do the semantic indices align without a single

1013
00:50:04,909 --> 00:50:08,849
clash? If every single constraint is internally

1014
00:50:08,849 --> 00:50:11,110
consistent, the utterance passes through the

1015
00:50:11,110 --> 00:50:13,269
filter and is declared a grammatical sentence.

1016
00:50:13,929 --> 00:50:16,429
If even a single feature clashes, it bounces

1017
00:50:16,429 --> 00:50:20,550
off. It's rejected. It's pure instantaneous constraint

1018
00:50:20,550 --> 00:50:24,710
satisfaction. But, um, I do have one major pushback

1019
00:50:24,710 --> 00:50:27,670
on this. Let's hear it. If we are just checking

1020
00:50:27,670 --> 00:50:30,309
mother nodes against daughter nodes locally within

1021
00:50:30,309 --> 00:50:33,269
these construction blocks, doesn't that severely

1022
00:50:33,269 --> 00:50:37,010
limit the grammar? How do we handle massive overarching

1023
00:50:37,010 --> 00:50:39,829
rules that govern the whole sentence? The textbook

1024
00:50:39,829 --> 00:50:42,070
has a whole section on locality that caught my

1025
00:50:42,070 --> 00:50:44,750
eye. You've hit on the exact limitation of the

1026
00:50:44,750 --> 00:50:47,030
model, but the authors argue this limitation

1027
00:50:47,030 --> 00:50:49,889
is actually its greatest triumph. A triumph?

1028
00:50:50,230 --> 00:50:54,010
How? Human languages are inherently local. A

1029
00:50:54,010 --> 00:50:56,469
verb at the very top of a complex sentence cannot

1030
00:50:56,469 --> 00:50:59,690
biologically select for a deeply buried obscure

1031
00:50:59,690 --> 00:51:02,309
property of a noun tucked inside a prepositional

1032
00:51:02,309 --> 00:51:05,190
phrase three levels down. The human brain simply

1033
00:51:05,190 --> 00:51:08,429
cannot track dependencies that convoluted. Our

1034
00:51:08,429 --> 00:51:11,289
working memory would buffer and crash. Exactly.

1035
00:51:11,630 --> 00:51:14,050
By forcing the grammatical math to operate strictly

1036
00:51:14,050 --> 00:51:16,289
within these local mother -daughter construction

1037
00:51:16,289 --> 00:51:19,849
blocks, the theory perfectly mirrors human cognitive

1038
00:51:19,849 --> 00:51:22,929
limitations. Oh, I see. It predicts that human

1039
00:51:22,929 --> 00:51:24,849
language will only ever reach a certain level

1040
00:51:24,849 --> 00:51:27,630
of complexity before it breaks down. The math

1041
00:51:27,630 --> 00:51:30,650
is constrained in the exact same way human memory

1042
00:51:30,650 --> 00:51:35,000
is constrained. The ultimate synthesis. The mathematics

1043
00:51:35,000 --> 00:51:37,820
perfectly echo the limits of the flesh. It is

1044
00:51:37,820 --> 00:51:39,800
truly a master class in scientific modeling.

1045
00:51:40,809 --> 00:51:44,139
It is. Well, we have covered a monumental amount

1046
00:51:44,139 --> 00:51:46,699
of ground today. We certainly have. We started

1047
00:51:46,699 --> 00:51:49,079
by stripping away the absurdity of prescriptive

1048
00:51:49,079 --> 00:51:51,739
dress codes, we learned how to draw context -free

1049
00:51:51,739 --> 00:51:54,400
trees, and then we shattered them to look at

1050
00:51:54,400 --> 00:51:57,159
the DNA of words, the complex feature structures

1051
00:51:57,159 --> 00:51:59,519
and chemical valences that drive our speech.

1052
00:52:00,800 --> 00:52:03,159
We watched semantic indexes, cast the roles of

1053
00:52:03,159 --> 00:52:05,820
actors, traced mathematical gaps up through the

1054
00:52:05,820 --> 00:52:08,380
branches of questions, and finally arrived at

1055
00:52:08,380 --> 00:52:10,739
a unified theory of sign -based construction

1056
00:52:10,739 --> 00:52:13,480
grammar that perfectly mirrors the parallel processing

1057
00:52:13,480 --> 00:52:15,789
of the human mind. And I hope the relevance of

1058
00:52:15,789 --> 00:52:17,809
this deep dive is clear to anyone listening.

1059
00:52:18,530 --> 00:52:20,829
You are executing billions of these flawless

1060
00:52:20,829 --> 00:52:23,230
complex constraint satisfaction calculations

1061
00:52:23,230 --> 00:52:25,929
every single time you chat with a friend or order

1062
00:52:25,929 --> 00:52:29,409
a coffee. Every single day. But beyond understanding

1063
00:52:29,409 --> 00:52:32,050
yourself, this invisible architecture we map

1064
00:52:32,050 --> 00:52:34,309
today is the literal blueprint for the future.

1065
00:52:34,889 --> 00:52:37,409
It is the exact logic we are using to teach machines

1066
00:52:37,409 --> 00:52:40,190
how to parse, think, and communicate in the modern

1067
00:52:40,190 --> 00:52:43,090
world. But, as always on the Deep Dive, we don't

1068
00:52:43,090 --> 00:52:45,170
want to leave you with everything neatly tied

1069
00:52:45,170 --> 00:52:48,150
up in a bow. We want to leave you with a provocative

1070
00:52:48,150 --> 00:52:50,909
final thought, something to twist your brain

1071
00:52:50,909 --> 00:52:53,710
a little bit as you go about your day. In Appendix

1072
00:52:53,710 --> 00:52:56,269
B of the textbook, after all this rigorous work,

1073
00:52:56,829 --> 00:52:59,909
the authors briefly nod to rival syntactic theories.

1074
00:53:00,329 --> 00:53:02,989
Rival theories. And one of them is called dependency

1075
00:53:02,989 --> 00:53:06,210
grammar. Dependency grammar makes a radical proposition.

1076
00:53:06,849 --> 00:53:10,570
It suggests that phrases Noun phrases, verb phrases,

1077
00:53:11,050 --> 00:53:13,030
the very branches of the structural trees we

1078
00:53:13,030 --> 00:53:16,030
just spent an hour painstakingly building, don't

1079
00:53:16,030 --> 00:53:19,639
actually exist in the brain at all. Wait. No

1080
00:53:19,639 --> 00:53:23,110
phrases? But the trees, we need the trees to

1081
00:53:23,110 --> 00:53:25,829
resolve ambiguity. Dependency grammar argues

1082
00:53:25,829 --> 00:53:28,150
that there are only direct point -to -point,

1083
00:53:28,369 --> 00:53:30,809
word -to -word dependencies. The verb points

1084
00:53:30,809 --> 00:53:33,349
an invisible mathematical arrow directly to the

1085
00:53:33,349 --> 00:53:36,269
noun. No hierarchical phrase branches required.

1086
00:53:36,929 --> 00:53:40,460
That... That feels like heresy after everything

1087
00:53:40,460 --> 00:53:43,179
we just discussed. It is a radical departure.

1088
00:53:43,440 --> 00:53:46,550
But if our brains really are just evaluating

1089
00:53:46,550 --> 00:53:49,610
local semantic constraints incrementally, instantly,

1090
00:53:49,789 --> 00:53:51,630
word by word, as the psycholinguistics chapter

1091
00:53:51,630 --> 00:53:55,420
proved, it raises the ultimate question. What

1092
00:53:55,420 --> 00:53:58,260
if the entire beautiful architecture of phrase

1093
00:53:58,260 --> 00:54:01,699
structure trees we just built is merely a useful

1094
00:54:01,699 --> 00:54:05,010
hallucination? Wow. What if it's just a diagram

1095
00:54:05,010 --> 00:54:07,530
linguists invented to give physical shape to

1096
00:54:07,530 --> 00:54:10,210
something that is entirely formless? What if

1097
00:54:10,210 --> 00:54:13,489
the human brain bypasses the tree entirely, mapping

1098
00:54:13,489 --> 00:54:16,650
words directly to words in a pure, unbranched

1099
00:54:16,650 --> 00:54:20,630
web of instant dependencies? That completely

1100
00:54:20,630 --> 00:54:22,510
shatters the anatomy metaphor we started with.

1101
00:54:22,630 --> 00:54:25,369
Doesn't it? It suggests that when we look at

1102
00:54:25,369 --> 00:54:27,650
the invisible architecture of language, the bones

1103
00:54:27,650 --> 00:54:30,409
might not even be there. It might just be pure

1104
00:54:30,409 --> 00:54:32,809
energy, pure connection. A fascinating thought

1105
00:54:32,809 --> 00:54:35,690
to leave on. I highly recommend you mull that

1106
00:54:35,690 --> 00:54:38,050
over the next time you effortlessly disambiguate

1107
00:54:38,050 --> 00:54:39,710
a sentence. See you next time.